An Introduction to LP Relaxations for MAP Inference

Transcription

1 An Introduction to LP Relaxations for MAP Inference Adrian Weller MLSALT4 Lecture Feb 27, 2017 With thanks to David Sontag (NYU) for use of some of his slides and illustrations For more information, see 1 / 39

2 High level overview of our 3 lectures 1. Directed and undirected graphical models (last Friday) 2. LP relaxations for MAP inference (today) 3. Junction tree algorithm for exact inference, belief propagation, variational methods for approximate inference (this Friday) Further reading / viewing: Murphy, Machine Learning: a Probabilistic Perspective Barber, Bayesian Reasoning and Machine Learning Bishop, Pattern Recognition and Machine Learning Koller and Friedman, Probabilistic Graphical Models Wainwright and Jordan, Graphical Models, Exponential Families, and Variational Inference 2 / 39

3 Example of MAP inference: image denoising Inference is combining prior beliefs with observed evidence to form a prediction. MAP inference 3 / 39

4 Example of MAP inference: protein side-chain placement Probabilistic inference Find minimum energy configuration of amino acid side-chains Given along desired a fixed 3D carbon structure, backbone: choose amino-acids giving the most stable folding (Yanover, Meltzer, Weiss 06) Side-chain X 1 (corresponding to 1 amino acid) X 3 X 2 θ 13 (x 1, x 3 ) X 3 X 1 Potential function for each edge θ 12 (x 1, x 2 ) a table X 2 Protein backbone θ 34 (x 3, x 4 ) Joint distribution over the variables is given by Orientations of the side-chains are represented by discretized angles called rotamers Rotamer choices for nearby amino acids are energetically coupled Partition function Key problems (attractive and repulsive forces) Find marginals: Find most likely assignment (MAP): Focus of this talk X 4 4 / 39

5 Outline of talk Background on undirected graphical models Basic LP relaxation Tighter relaxations Message passing and dual decomposition We ll comment on When is an LP relaxation tight 5 / 39

6 Background: undirected graphical models Powerful way to represent relationships across variables Many applications including: computer vision, social network analysis, deep belief networks, protein folding... In this talk, focus on pairwise models with discrete variables (sometimes binary) Example: Grid for computer vision 6 / 39

7 Background: undirected graphical models Discrete variables X 1,..., X n with X i {0,..., k i 1} Potential functions, will somehow write as vector θ Write x = (... x 1,..., x n,... ) for one overcomplete configuration of all variables, θ x for its total score Probability distribution given by p(x) = 1 Z exp(θ x) To ensure probabilities sum to 1, need normalizing constant or partition function Z = x exp (θ x) We are interested in maximum a posteriori (MAP) inference i.e., find a global configuration with highest probability x arg max p(x) = arg max θ x 7 / 39

8 Background: how do we write potentials as a vector θ? θ x means the total score of a configuration x, where we sum over all potential functions If we have potential functions θ c over some subsets c C of variables, then we want c C θ c(x c ), where x c means a configuration of variables just in the subset c θ c (x c ) provides a measure of local compatibility, a table of values 8 / 39

9 Background: how do we write potentials as a vector θ? θ x means the total score of a configuration x, where we sum over all potential functions If we have potential functions θ c over some subsets c C of variables, then we want c C θ c(x c ), where x c means a configuration of variables just in the subset c θ c (x c ) provides a measure of local compatibility, a table of values If we only have some unary/singleton potentials θ i and edge/pairwise potentials θ ij then we can write the total score as θ i (x i ) + θ ij (x i, x j ) (i,j) i Indices? Usually assume either no unary potentials (absorb them into edges) or one for every variable, leading to a graph topology (V, E) with total score θ i (x i ) + θ ij (x i, x j ) i V ={1,...,n} (i,j) E 8 / 39

10 Background: overcomplete representation The overcomplete representation conveniently allows us to write θ x = θ i (x i ) + θ ij (x i, x j ) i V (i,j) E Concatenate singleton and edge terms into big vectors θ i (0) 1[X i = 0] θ i (1) 1[X i = 1] θ =... θ ij (0, 0) x =... 1[X i = 0, X j = 0] θ ij (0, 1) 1[X i = 0, X j = 1] θ ij (1, 0) 1[X i = 1, X j = 0] θ ij (1, 1) 1[X i = 1, X j = 1] There are many possible values of x how many? 9 / 39

11 Background: overcomplete representation The overcomplete representation conveniently allows us to write θ x = θ i (x i ) + θ ij (x i, x j ) i V (i,j) E Concatenate singleton and edge terms into big vectors θ i (0) 1[X i = 0] θ i (1) 1[X i = 1] θ =... θ ij (0, 0) x =... 1[X i = 0, X j = 0] θ ij (0, 1) 1[X i = 0, X j = 1] θ ij (1, 0) 1[X i = 1, X j = 0] θ ij (1, 1) 1[X i = 1, X j = 1] There are many possible values of x how many? i V k i 9 / 39

12 Background: Binary pairwise models θ x is the score of a configuration x Probability distribution given by p(x) = 1 Z exp(θ x) For MAP inference, want x arg max p(x) = arg max θ x Want to optimize over {0, 1} coordinates of overcomplete configuration space corresponding to all 2 n possible settings The convex hull of these defines the marginal polytope M Each point µ M corresponds to a probability distribution over the 2 n configurations, giving a vector of marginals 10 / 39

13 Background: the marginal polytope (all valid marginals) µ = 0! 1! 0! 1! 1! 0! 0" 0" 1" 0" 0! 0! 0! 1! 0! 0! 1! 0" Marginal polytope! (Wainwright & Jordan, 03)! µ = 1 µ 2 + µ valid marginal probabilities! X 1! = 1! 1! 0! 0! 1! 1! 0! 1" 0" 0" 0" 0! 1! 0! 0! 0! 0! 1! 0" Assignment for X 1" Assignment for X 2" Assignment for X 3! Edge assignment for" X 1 X 3! Edge assignment for" X 1 X 2! Edge assignment for" X 2 X 3! X 1! = 0! X 2! = 1! X 3 =! 0! X 2! = 1! X 3 =! 0! Figure 2-1: Illustration of the marginal polytope for a Markov random field with three nodes 11 / 39

14 Background: overcomplete and minimal representations The overcomplete representation is highly redundant, e.g. µ i (0) + µ i (1) = 1 i How many dimensions if n binary variables with m edges? 12 / 39

15 Background: overcomplete and minimal representations The overcomplete representation is highly redundant, e.g. µ i (0) + µ i (1) = 1 i How many dimensions if n binary variables with m edges? 2n + 4m Instead, we sometimes pick a minimal representation What s the minimum number of dimensions we need? 12 / 39

16 Background: overcomplete and minimal representations The overcomplete representation is highly redundant, e.g. µ i (0) + µ i (1) = 1 i How many dimensions if n binary variables with m edges? 2n + 4m Instead, we sometimes pick a minimal representation What s the minimum number of dimensions we need? n + m For example, we could use q = (q 1,..., q n,..., q ij,... ) where q i = µ i (1) i, q ij = µ ij (1, 1) (i, j), then ( ) ( ) ( ) 1 qi 1 qj 1 + qij q µ i =, µ q j =, µ i q ij = i q j q j q ij j q i q ij q ij Note many other possible minimal representations 12 / 39

17 LP relaxation: MAP as an integer linear program (ILP) MAP inference as a discrete optimization problem is to identify a configuration with maximum total score x arg max θ i (x i ) + θ ij (x i, x j ) x i V ij E = arg max x = arg max µ θ x θ i (x i )µ i (x i ) + θ ij (x i, x j )µ ij (x i, x j ) x i x i,x j i V = arg max θ µ µ Any other constraints? ij E s.t. µ is integral 13 / 39

18 What are the constraints? Force every cluster of variables to choose a local assignment: µ i (x i ) {0, 1} i V, x i x i µ i (x i ) = 1 i V µ ij (x i, x j ) {0, 1} ij E, x i, x j x i,x j µ ij (x i, x j ) = 1 ij E Enforce that these assignments are consistent: µ i (x i ) = x j µ ij (x i, x j ) ij E, x i µ j (x j ) = x i µ ij (x i, x j ) ij E, x j 14 / 39

19 MAP as an integer linear program (ILP) subject to: MAP(θ) = max θ i (x i )µ i (x i ) + θ ij (x i, x j )µ ij (x i, x j ) µ x i x i,x j i V = max µ θ µ ij E µ i (x i ) {0, 1} i V, x i (edge terms?) x i µ i (x i ) = 1 i V µ i (x i ) = x j µ ij (x i, x j ) ij E, x i µ j (x j ) = x i µ ij (x i, x j ) ij E, x j Many good off-the-shelf solvers, such as CPLEX and Gurobi 15 / 39

20 Linear programming (LP) relaxation for MAP Integer linear program was: subject to MAP(θ) = max µ θ µ µ i (x i ) {0, 1} i V, x i x i µ i (x i ) = 1 i V µ i (x i ) = x j µ ij (x i, x j ) ij E, x i µ j (x j ) = x i µ ij (x i, x j ) ij E, x j Now relax integrality constraints, allow variables to be between 0 and 1: µ i (x i ) [0, 1] i V, x i 16 / 39

21 Basic LP relaxation for MAP LP(θ) = max µ θ µ s.t. µ i (x i ) [0, 1] i V, x i x i µ i (x i ) = 1 i V µ i (x i ) = x j µ ij (x i, x j ) ij E, x i µ j (x j ) = x i µ ij (x i, x j ) ij E, x j Linear programs can be solved efficiently: simplex, interior point, ellipsoid algorithm Since the LP relaxation maximizes over a larger set, its value can only be higher MAP(θ) LP(θ) 17 / 39

22 The local polytope LP(θ) = max µ θ µ s.t. µ i (x i ) [0, 1] i V, x i x i µ i (x i ) = 1 i V µ i (x i ) = µ j (x j ) = µ ij (x i, x j ) ij E, x i x i µ ij (x i, x j ) ij E, x j All these constraints are linear Hence define a polytope in the space of marginals Here we enforced only local (pairwise) consistency, which defines the local polytope If instead we had optimized over the marginal polytope, which enforces global consistency, then we would have MAP(θ) =LP(θ), i.e. the LP is tight why? why don t we do this? 18 / 39

23 Tighter relaxations of the marginal polytope Enforcing consistency of pairs of variables leads to the local polytope L 2 The marginal polytope enforces consistency over all variables M = L n Natural to consider the Sherali-Adams hierarchy of successively tighter relaxations L r 2 r n which enforce consistency over clusters of r variables Just up from the local polytope is the triplet polytope TRI= L 3 19 / 39

24 Stylized illustration of polytopes marginal polytope M = L n global consistency... triplet polytope L 3 triplet consistency local polytope L 2 pair consistency More accurate Less accurate More computationally intensive Less computationally intensive 20 / 39

25 Stylized illustration of polytopes marginal polytope M = L n global consistency... triplet polytope L 3 triplet consistency local polytope L 2 pair consistency More accurate Less accurate More computationally intensive Less computationally intensive Can be shown that for binary variables, TRI=CYC, the cycle polytope, which enforces consistency over all cycles In general, TRI CYC, open problem if TRI = CYC [SonPhD 3] 20 / 39

26 When is the LP tight? For a model without cycles, local polytope L 2 =M marginal polytope, hence the basic LP ( first order ) is always tight More generally, if a model has treewidth r then LP+L r+1 is tight [WaiJor04] STRUCTURE Separately, if we allow any structure but restrict the class of potential functions, interesting results are known POTENTIALS For example, the basic LP is tight if all potentials are supermodular Fascinating recent work [KolThaZiv15]: if we do not restrict structure, then for any given family of potentials, either the basic LP relaxation is tight or the problem class is NP-hard! Identifying HYBRID conditions is an exciting current research area 21 / 39

27 When is MAP inference (relatively) easy? Tree Attractive (binary) model STRUCTURE POTENTIALS Both can be handled efficiently by the basic LP relaxation, LP+L 2 22 / 39

28 Cutting planes µ max θ µ µ max θ µ µ max θ µ µ max θ µ θ θ θ θ µ µ µ µ (a) (b) (c) (d) Figure 2-6: Illustration of the cutting-plane algorithm. (a) Solve the LP relaxation. (b) Find a violated constraint, add it to the relaxation, and repeat. (c) Result of solving the tighter LP relaxation. (d) Finally, we find the MAP assignment. of the clusters considered. Even optimizing over TRI(G), the first lifting of the Sherali- SonPhD Adams hierarchy, is impractical for all but the smallest of graphical models. Our approach is motivated by the observation that it may not be necessary to add all of the constraints that make up a higher-order relaxation such as TRI(G). In particular, it possible that the pairwise LP relaxation alone is close to being tight, and that only a few carefully chosen 23 / 39

29 Cutting planes Invalid! constraint" µ max θ µ µ θ Useless! constraint" want to choose constraints that are both valid and useful. A v SonPhD es not cut off any of the integer points. One way to guarantee We want to add constraints that are both valid and useful are useful is to use them within the cutting-plane algorithm. C Valid: does not cut off any integer points separate the current fractional solution from the rest of the int Useful: leads us to update to a better solution tightest valid constraints to add are the facets of the margina 24 / 39

30 Methods for solving general integer linear programs Local search Start from an arbitrary assignment (e.g., random). Choose a variable. Branch-and-bound Exhaustive search over space of assignments, pruning branches that can be provably shown not to contain a MAP assignment Can use the LP relaxation or its dual to obtain upper bounds Lower bound obtained from value of any assignment found Branch-and-cut (most powerful method; used by CPLEX & Gurobi) Same as branch-and-bound; spend more time getting tighter bounds Adds cutting-planes to cut off fractional solutions of the LP relaxation, making the upper bound tighter 25 / 39

31 Message passing Can be a computationally efficient way to obtain or approximate a MAP solution, takes advantage of the graph structure Classic example is max-product belief propagation (BP) Sufficient conditions are known s.t. this will always converge to the solution of the basic LP, includes that the basic LP is tight [ParkShin-UAI15] In general, however, this may not converge to the LP solution (even for supermodular potentials) Other methods have been developed, many relate to dual decomposition / 39

32 Dual decomposition and reparameterizations Consider the MAP problem for pairwise Markov random fields: MAP(θ) = max θ i (x i ) + θ ij (x i, x j ). x i V ij E If we push the maximizations inside the sums, the value can only increase: MAP(θ) max θ i (x i ) + max θ ij (x i, x j ) x i x i,x j i V ij E Note that the right-hand side can be easily evaluated One can always reparameterize a distribution by operations like θi new (x i ) = θi old (x i ) + f (x i ) θij new (x i, x j ) = θij old (x i, x j ) f (x i ) for any function f (x i ), without changing the distribution/energy 27 / 39

33 Dual decomposition Introduction to Dual Decomposition for Inference f (x 1,x 2 ) x 1 f1(x 1) x 2 f2(x 2) f(x1,x2) x 1 x 2 g(x1,x3) h(x2,x4) x 3 x 4 k(x3,x4) f1(x 1) f2(x 2) x 1 + x 1 x 2 + x g (x 1,x 3 ) 2 g1(x 1 ) h2(x 2 ) g3(x 3 ) h (x 2,x 4 ) g1(x 1 ) h2(x 2 ) x g3(x 3 ) k4(x 4 ) 3 h4(x 4 ) + x 3 x 4 + x 4 k3(x 3 ) h4(x 4 ) x 3 x 4 k (x 3,x 4 ) k3(x 3 ) k4(x 4 ) Figure 1.2: Illustration of the the dual decomposition objective. Left: The original pairwise model consisting of four factors. Right: The maximization 28 / 39

34 Dual decomposition Define: θ i (x i ) = θ i (x i ) + ij E δ j i (x i ) θ ij (x i, x j ) = θ ij (x i, x j ) δ j i (x i ) δ i j (x j ) It is easy to verify that θ i (x i ) + θ ij (x i, x j ) = i ij E i Thus, we have that: θ i (x i ) + ij E θ ij (x i, x j ) x MAP(θ) = MAP( θ) max θi (x i ) + max θij (x i, x j ) x i x i,x j i V ij E Every value of δ gives a different upper bound on the value of the MAP The tightest upper bound can be obtained by minimizing the RHS with respect to δ 29 / 39

35 Dual decomposition We obtain the following dual objective: L(δ) = ( max θ i (x i ) + ) δ j i (x i ) + ( ) max θ ij (x i, x j ) δ j i (x i ) δ i j (x j ), x i x i,x j i V ij E ij E DUAL-LP(θ) = min L(δ) δ This provides an upper bound on the MAP assignment MAP(θ) DUAL-LP(θ) L(δ) How can find δ which give tight bounds? 30 / 39

36 Solving the dual efficiently Many ways to solve the dual linear program, i.e. minimize with respect to δ: ( max θ i (x i ) + ) δ j i (x i ) + ( ) max θ ij (x i, x j ) δ j i (x i ) δ i j (x j ), x i x i,x j i V ij E ij E One option is to use the subgradient method Can also solve using block coordinate-descent, which gives algorithms that look very much like max-sum belief propagation 31 / 39

37 Solving the dual efficiently Many ways to solve the dual linear program, i.e. minimize with respect to δ: ( max θ i (x i ) + ) δ j i (x i ) + ( ) max θ ij (x i, x j ) δ j i (x i ) δ i j (x j ), x i x i,x j i V ij E ij E One option is to use the subgradient method Can also solve using block coordinate-descent, which gives algorithms that look very much like max-sum belief propagation 31 / 39

38 Solving the dual efficiently Many ways to solve the dual linear program, i.e. minimize with respect to δ: ( max θ i (x i ) + ) δ j i (x i ) + ( ) max θ ij (x i, x j ) δ j i (x i ) δ i j (x j ), x i x i,x j i V ij E ij E One option is to use the subgradient method Can also solve using block coordinate-descent, which gives algorithms that look very much like max-sum belief propagation 31 / 39

39 Max-product linear programming (MPLP) algorithm Input: A set of potentials θ i (x i ), θ ij (x i, x j ) Output: An assignment x 1,..., x n that approximates a MAP solution Algorithm: Initialize δ i j (x j ) = 0, δ j i (x i ) = 0, ij E, x i, x j Iterate until small enough change in L(δ): For each edge ij E (sequentially), perform the updates: δ j i (x i ) = 1 2 δ j i (x i ) + 1 [ ] 2 max θ ij (x i, x j ) + δ i x j (x j ) j δ i j (x j ) = 1 2 δ i j (x j ) + 1 [ ] 2 max θ ij (x i, x j ) + δ j x i (x i ) i x i x j where δ j i (x i ) = θ i (x i ) + ik E,k j δ k i(x i ) Return x i arg maxˆxi θ δ i (ˆx i) 32 / 39

40 Generalization to arbitrary factor graphs [SonGloJaa11] Introduction to Dual Decomposition for Inference Inputs: A set of factors θ i(x i), θ f (x f ). Output: An assignment x 1,...,x n that approximates the MAP. Algorithm: Initialize δ fi(x i)=0, f F, i f,x i. Iterate until small enough change in L(δ) (seeeq.1.2): For each f F, perform the updates δ fi(x i)= δ f i (x 1 i)+ f max θ f (x f )+ x f\i î f δ f î (xî), (1.16) simultaneously for all i f and x i.wedefineδ f i (x i)=θ i(x i)+ ˆf=f δ ˆfi (x i). Return x i arg maxˆxi θδ i (ˆx i)(seeeq.1.6). Figure 1.4: Description of the MPLP block coordinate descent algorithm for minimizing the dual L(δ) (see Section 1.5.2). Similar algorithms can be devised for different choices of coordinate blocks. See sections and The assignment returned in the final step follows the decoding scheme 33 / 39

41 Experimental results Performance on stereo vision inference task: Solved optimally! Duality gap! Dual obj.! Objective! Decoded assignment! Iteration! 34 / 39

42 Dual decomposition = basic LP relaxation Recall we obtained the following dual linear program: L(δ) = ( max θ i (x i ) + ) δ j i (x i ) + ( ) max θ ij (x i, x j ) δ j i (x i ) δ i j (x j ), x i x i,x j i V ij E ij E DUAL-LP(θ) = min L(δ) δ We showed two ways of upper bounding the value of the MAP assignment: MAP(θ) Basic LP(θ) (1) MAP(θ) DUAL-LP(θ) L(δ) (2) Although we derived these linear programs in seemingly very different ways, in turns out that: Basic LP(θ) = DUAL-LP(θ) [SonGloJaa11] The dual LP allows us to upper bound the value of the MAP assignment without solving an LP to optimality 35 / 39

43 Linear programming duality MAP assignment! µ x*! (Dual) LP relaxation! (Primal) LP relaxation! Integer linear program! MAP(θ) Basic LP(θ) = DUAL-LP(θ) L(δ) 36 / 39

44 Cutting-plane algorithm µ max θ µ µ max θ µ µ max θ µ µ max θ µ θ θ θ θ µ µ µ µ (a) (b) (c) (d) Figure 2-6: Illustration of the cutting-plane algorithm. (a) Solve the LP relaxation. (b) Find a violated constraint, add it to the relaxation, and repeat. (c) Result of solving the tighter LP relaxation. (d) Finally, we find the MAP assignment. is motivated by the observation that it may not be necessary to add all of the constraints that make up a higher-order relaxation such as TRI(G). In particular, it possible that the pairwise LP relaxation alone is close to being tight, and that only a few carefully chosen constraints would suffice to obtain an integer solution. Our algorithms tighten the relaxation in a problem-specific way, using additional computation just for the hard parts of each instance. We illustrate the general approach in 37 / 39

45 Conclusion LP relaxations yield a powerful approach for MAP inference Naturally lead to considerations of polytope or cutting planes dual decomposition and message passing Close relationship to methods for marginal inference Help build understanding as well as develop new algorithmic tools Exciting current research Thank you 38 / 39

46 References V. Kolmogorov, J. Thapper, and S. Živný. The power of linear programming for general-valued CSPs. SIAM Journal on Computing, 44(1):136, S. Park and J. Shin. Max-product belief propagation for linear programming: applications to combinatorial optimization. In UAI, D. Sontag. Approximate inference in graphical models using LP relaxations. PhD thesis, MIT, D. Sontag, A. Globerson, and T. Jaakkola. Introduction to dual decomposition for inference. In Optimization for Machine Learning, MIT Press, J. Thapper and S. Živný. The complexity of finite-valued CSPs. arxiv technical report, M. Wainwright and M. Jordan. Treewidth-based conditions for exactness of the Sherali-Adams and Lasserre relaxations. Univ. California, Berkeley, Technical Report, 671:4, M. Wainwright and M. Jordan. Graphical models, exponential families and variational inference. Foundations and Trends in Machine Learning, 1(1-2):1305, / 39